Ayman Mahfuz

My research focuses on leveraging LLM-based multiagent interactions to enhance reasoning and decision-making in the medical domain. I design Python systems to evaluate the consistency of AI models in medical diagnoses, analyzing how contextual variations influence outcomes. This work combines multiagent collaboration and reasoning with cutting-edge AI techniques to improve reliability in medical decision-making.

• Led a team in the systematic design and implementation of a comprehensive Python script for assessing the consistency of MedAgents in diagnosing medical conditions. Utilized the Autogen library and GPT-3.5 API, incorporating features such as exponential backoff for API rate limiting, offline script execution, and organized data collection. Ensured robust performance through well-structured code, modular design, and extensive documentation.
• Engineered robust data pipelines using Pandas and NumPy for data preprocessing and cleaning. Developed methods to introduce variations to medical questions, ensuring clean and structured datasets for analysis. Managed data storage and processing in JSON format, facilitating efficient and accurate data handling.
• Developed models to evaluate the impact of contextual changes on the diagnostic consistency of MedAgents. Assessed the influence of variations such as patient age or symptom modifications on model responses, identifying patterns and key factors affecting diagnostic outcomes. Utilized advanced machine learning techniques to analyze and interpret results, contributing to the project's goal of enhancing the reliability of AI in medical question answering.
• Conducted advanced data analysis and visualization using a variety of statistical and machine learning techniques. Employed tools such as Scikit-learn, Statsmodels, and Matplotlib for analyses including descriptive statistics, correlation matrices, ANOVA, Chi-square tests, Cohen's Kappa, and logistic regression. Created comprehensive dashboards and visualizations to track progress, identify patterns, and derive actionable insights, significantly contributing to the project's overall success.

I conduct media research by studying how people interact with news, platforms, and each other, by designing systems and using machine learning to evaluate political opinions, storytelling patterns, and societal divides. My contributions include building a 150-million-entry dataset, developing DistilBERT models for large-scale analysis, and designing the full system architecture for MTurk-integrated React games to track and analyze user behavior.

• Engineered robust Python pipelines for scraping, preprocessing, and uploading data to BigQuery. Employed various scraping techniques (APIs, sitemaps, HTML parsing), leveraged Pandas and NumPy heavily, and ensured data integrity with automated cron jobs for daily updates. Processed over 50 million news articles and 70 million comments from 12+ domains, maintaining a comprehensive database dating back to 1851.
• Designed and deployed a research platform using React, Tailwind, Flask, and Firebase, featuring 3 interactive games (Higher/Lower, Slider, and Timeline) with MTurk integration, real-time analytics tracking 15+ metrics, and 99.9% uptime. The platform includes a modular Flask backend that manages random game assignments, worker ID tracking, and question set management, seamlessly integrated with Firebase for data logging and storage.
• Developed dynamic dashboards and visualizations using SQL, Pandas, Matplotlib, and Looker Studio to track data collection progress, identify and programmatically fill data gaps, enhancing the database's reliability and accessibility for research purposes.
• Led machine learning initiatives, creating and evaluating models to classify news headlines and comments. Fine-tuned a DistilBERT model using Hugging Face, achieving high accuracy and performance metrics (precision, recall, F1 score) through comprehensive evaluations.
• Conducted advanced research projects: analyzed clickbait trends over a decade and investigated the prevalence of personal stories in comments. Leveraged machine learning algorithms and extensive data analysis to derive actionable insights, contributing to upcoming research publications that aim to understand and combat the spread of misinformation online.

I use the latest transformer technology to advance pancreas and organ segmentation, applying state-of-the-art models like MedSAM2 to preprocess 3D MRI data. This work aims to identify early indicators of diabetes and pancreatic cancer, enabling earlier and more accurate diagnoses for patients.

• Led a team of 3 members to revolutionize abdominal organ segmentation by developing advanced machine learning models for medical imaging. Built on existing MONAI frameworks to segment the pancreas from MRI scans, enhancing accuracy and performance using the latest vision transformer technologies and the TransUNet library.
• Engineered robust data preprocessing pipelines: Designed and implemented Python scripts for preprocessing 3D MRI images, including clipping, normalization, and extraction of 2D slices, using libraries such as Nibabel, Pydicom, NumPy, and H5py. Ensured efficient data handling and storage for large-scale medical datasets, facilitating downstream analysis and model training.
• Developed and fine-tuned machine learning models for pancreas segmentation using advanced architectures like TransUNet and ResNet-50 with ViT. Implemented model training, validation, and evaluation processes using PyTorch, achieving significant improvements over existing benchmarks. Utilized metrics such as Dice Score to assess model performance.
• Conducted comprehensive data analysis and visualization: Employed statistical and machine learning techniques to analyze the impact of contextual changes on model performance. Utilized tools such as Scikit-learn, Statsmodels, and Matplotlib to perform descriptive statistics, correlation analyses, and visualize segmentation results, driving insights into model behavior and robustness.

I’m building an NLP-based chatbot to engage news readers and analyzing linguistic patterns to enhance the interaction. My contributions focus on Python scripting and publishing insights at CHI 2024.

• Project: Towards Designing a Question-Answering Chatbot for Online News
• Led the development of an NLP-driven chatbot to augment online news reader engagement, employing deep learning and AI techniques. Directed comprehensive studies and analyses, culminating in findings on human-chatbot interaction dynamics.
• Executed sophisticated text analytics and data labeling using Python, encompassing Parts of Speech Tagging, LIWC, and clustering on sentence embeddings, to derive intricate linguistic patterns and insights.
• Collaborated with a cross-disciplinary team of professors and graduate students, driving content creation and ensuring methodological precision. Contributed core analytical insights and visualizations to the research paper published at the CHI 2024 conference.
• Authored Python scripts for in-depth data analysis, generating insightful graphs and visualizations that formed the backbone of the research findings, illustrating complex human-chatbot interaction patterns.

At Lockheed Martin, I optimized CRM workflows, introduced RPA solutions, and cleaned up their Configuration Database to boost operational efficiency.

• Spearheaded the development and optimization of Customer Relationship Management (CRM) workflows at Lockheed Martin, achieving a centralized device data framework that enhanced enterprise operational efficiency.
• Engineered advanced CRM solutions by integrating JavaScript for flow enhancements and implementing Robotic Process Automation (RPA), streamlining the data de-duplication process and elevating data integrity.
• Administered and refined the Configuration Database, successfully purging redundant records and bolstering data accuracy. Synthesized and presented data-driven insights to executives, highlighting the tangible impact on operational efficiency and guiding strategic decisions.

As the youngest participant, I gained exposure to AI, business strategies, and professional development while collaborating on tech-focused initiatives with industry leaders.

• Actively engaged with industry leaders at the AT&T Summer Learning Academy, gaining insights into AI, personal development, and emerging business trends, thereby enriching my technological acumen and industry foresight.
• Collaborated on tech-focused portfolio strategies, contributing fresh perspectives as the sole high school participant among international undergraduates, which enhanced the diversity of thought in strategy sessions.
• Leveraged the unique opportunity to broaden my understanding of technology's integral role in shaping business strategies and societal trends, positioning myself at the forefront of future tech innovators.

I helped Austin’s post-COVID recovery by improving loan processing workflows for small businesses. My work included Python scripting and data visualization to streamline operations.

• Drove Austin's post-COVID economic recovery by innovating data analytics processes, accelerating business loan facilitation and enhancing operational efficiency for the small business relief program.
• Designed and implemented Excel algorithms and Python scripts, revolutionizing loan application processing and rectification, significantly reducing processing time and errors.
• Pioneered the use of Power BI for dynamic business data visualization, delivering actionable insights that informed strategic relief efforts. Recognized for exceptional performance with a one-month internship extension and entrusted as the lead for automating critical operational processes.
• Instrumental in bolstering the resilience of Austin's small business community during the pandemic, earning accolades for superior performance and enduring contributions to the city's economic recovery initiatives.